flapping airplanes

an exploration of AI training efficiency — the name comes from the analogy of early airplane designers who tried to build planes that flap their wings like birds. they were copying the surface behavior (flight) instead of the underlying principle (lift via airfoil). the question being asked: are current deep learning training methods "flapping" — imitating what works without understanding why — and are there fundamentally more efficient approaches waiting to be discovered?

the specific research directions this points at include: sparse training (activate fewer parameters per forward pass), continual/online learning (learn from a stream of experience rather than a fixed dataset), and biologically-inspired learning rules that don't require backprop. the hypothesis is that gradient descent on massive static datasets is like flapping — it works, but it may be a local optimum in the space of learning algorithms, not the global one.